Machine-learning-based prediction of key operational parameters in a CH 4 /H 2 /air swirl combustor from a flame chemiluminescence spectrum

Abstract The demand for replacing hydrocarbon fuel with hydrogen (H 2 ) in existing combustion systems has increased to achieve a carbon-neutral society, subsequently necessitating developments of advanced monitoring technologies for safer operations of combustion systems. This study proposes a method that effectively predicts key operational parameters—total combustion flow rate ( V ˙ ), fuel–air equivalence ratio ( ϕ ), and H 2 blend ratio ( X H 2 )—for a given combustor, directly from the chemiluminescence spectrum of a model low-swirl combustor. This method is based on the fact that the combustion field is governed by these key operational parameters, and significant information about the field can be contained in the chemiluminescence spectrum. Since it is difficult to obtain an analytical relationship between the spectrum and the operational parameters, a predictive model was developed in a data-driven manner. V ˙ , ϕ , and X H 2 were varied between 80–140 l min −1 , 0.7–1.0, and 0–30 mol%, respectively, resulting in 441 experimental conditions, with 500 spectra collected for each condition. Specifically, the model consisted of a convolutional autoencoder along with separate regressors, as these predictions share the same spectral features. Once optimized, the final model was able to predict V ˙ , ϕ , and X H 2 within ±2.994 l min −1 , ±0.012, and ±2.252 mol%, respectively, with 96% probability. A gradient-weighted regression activation mapping analysis confirmed that the model employs all key features of OH*, CH*, C 2 *, CO 2 *, and H 2 O* intensities in predicting each parameter, while compensating for the effects of the others.